Hugging Face Blog
·
2 months ago
AWS released new GPU instance families—P6 with NVIDIA Blackwell B200 and B300 chips, and P6e-GB200 UltraServers with up to 72 GPUs in a single NVLink domain—designed to support foundation model training and inference across pre-training, post-training, and inference phases. The B200 GPU delivers 2.25 PFLOPS of dense BF16/FP16 Tensor throughput and 180 GB of HBM3e memory, while the B300 variant provides 288 GB of memory and support for up to 13.5 PFLOPS of FP4 operations. These systems integrate with open-source tools like PyTorch, Kubernetes, and Prometheus to reduce communication bottlenecks and improve scaling efficiency for large-scale distributed training workloads.
OpenAI Blog
·
2 months ago
ChatGPT's user base expanded significantly in early 2026, with particularly strong growth among older adults and more equitable distribution across genders. Users over 35 represented the fastest-growing demographic segment during the first quarter. This shift indicates AI tools are moving beyond early-adopter populations into broader demographic groups.
Import AI
·
2 months ago
Researchers at the Institute for Law & AI propose "radical optionality," a framework where governments invest now in regulatory tools and institutional capacity they may need to govern powerful AI systems in the future, rather than either regulating heavily now or not preparing at all. The approach recommends specific near-term investments including transparency requirements, whistleblower protections, government auditing capacity, and technical talent recruitment at agencies like the UK's AISI, with estimated costs modest relative to potential benefits. This avoids premature regulation while building flexibility to respond to different futures as AI development unfolds.
OpenAI Blog
·
2 months ago
Enterprises are moving AI initiatives beyond pilot projects by implementing trust frameworks, governance structures, and standardized workflows. Organizations that successfully scale AI report measurable productivity gains across departments when they establish clear data quality standards and oversight mechanisms. This shift from experimental deployments to systematic integration requires investment in organizational processes rather than technology alone.
OpenAI Blog
·
2 months ago
OpenAI launched a Campus Network program inviting student clubs to join a global community focused on AI education and engagement. Interested student organizations can submit an interest form to access AI tools, organize events, and collaborate with peers at other institutions. Participating clubs gain resources to integrate AI literacy into their campus activities and connect with a broader student network exploring artificial intelligence.
Allen Institute (AI2)
·
2 months ago
Artificial Analysis adopted IFBench, an AI2-developed benchmark accepted to NeurIPS 2025, to measure how well language models follow complex natural-language instructions like specific word counts, sentence structures, and keyword placement simultaneously. IFBench scores cluster sharply by model family, with Grok 4.20 leading at 82.9%, followed by Google Gemini and OpenAI GPT models, while Claude models score substantially lower despite ranking higher on Artificial Analysis's broader Intelligence Index. The benchmark remains useful for comparing models because instruction-following capability develops independently from other heavily optimized areas like coding and tool use, preventing saturation unlike most evaluations that become obsolete within six months.
OpenAI Blog
·
2 months ago
OpenAI created DeployCo, a new division focused on helping organizations deploy frontier AI systems into production environments. The company will work with enterprises to implement AI solutions that deliver measurable business results. This move allows OpenAI to directly support customers in translating AI capabilities into operational value rather than leaving deployment challenges to third parties.
Together AI
·
2 months ago
DeepSeek-V4's million-token context window relies on architectural changes using compressed sparse attention, heavily compressed attention, and sliding window attention that reduce key-value cache requirements. Together achieved 3.7M tokens of capacity on an NVIDIA HGX B200 node through cache management policies, compared to 1.2M without optimization. Serving V4 efficiently requires inference engines to handle multiple cache types, implement context-aware prefix caching policies, and choose endpoint configurations matched to workload characteristics—making long-context serving primarily a systems problem rather than just a model capability.
